The Effect of Estimation Error on Portfolio Weights

Using sample moments of asset returns as inputs to construct optimal portfolio weights results in volatile and extreme portfolio weights. This research therefore compares a number of methods that reduce the estimation error in the mean or in the covariance matrix. The methods are evaluated over different estimation periods with the help of simulated and empirical stock data, where the 1=N portfolio is used as a benchmark, because DeMiguel et al. (2009) found no portfolio that could significantly outperform the 1=N portfolio. In contrast to (DeMiguel et al., 2009), who used diversified equity portfolios to construct their portfolios, this paper uses individual stocks as available assets, which have higher idiosyncratic risk. It is found that the global minimum variance portfolios with short-sale constraints perform as good or better than the 1=N rule consistently across all datasets and estimation periods.